PermGen space outofmemory on slaves

We have experienced for some time now that Slaves die with PermGen space
OutOfMemoryError.

We are currently using Hudson 1.285, but it has been a problem with a number of
releases. Possibly all releases we have ever used - not entirely sure.

Our Master is restarted (hard via its Windows Service, without preparing it for
shutdown) at least once every day.

After a week or so, the Slaves (4 of them) start dropping like flies with
exhausted PermGen space.

All projects run via Ant. The builder setup looks like this:

<hudson.tasks.Ant>
<targets>ci.build ci.validate</targets>
<antOpts>-Xmx512m</antOpts>
<buildFile>build.ear.xml</buildFile>
<properties></properties>
</hudson.tasks.Ant>

I have changed the slaves to run with -XX:+HeapDumpOnOutOfMemoryError. I have
the dump available if anyone wants to have a look - it is 5MB zipped.

Looking at the dump with Eclipse's Memory Analyzer the prime suspects (according
to the tool - I'm a rookie on the subject) is:

46 instances of "hudson.remoting.RemoteClassLoader", loaded by
"sun.misc.Launcher$AppClassLoader @ 0x3007630" occupy 3.606.384 (60,70%) bytes.